Fuad Tabba <tabba@xxxxxxxxxx> writes: > Hi, > > As mentioned in the guest_memfd sync (2025-01-09), below is the state > diagram that uses the new states in this patch series, and how they > would interact with sharing/unsharing in pKVM: > > https://lpc.events/event/18/contributions/1758/attachments/1457/3699/Guestmemfd%20folio%20state%20page_type.pdf Thanks Fuad! I took a look at the state diagram [1] and the branch that this patch is on [2], and here's what I understand about the flow: 1. From state H in the state diagram, the guest can request to unshare a page. When KVM handles this unsharing, KVM marks the folio mappability as NONE (state J). 2. The transition from state J to state K or I is independent of KVM - userspace has to do this unmapping 3. On the next vcpu_run() from userspace, continuing from userspace's handling of the unshare request, guest_memfd will check and try to register a callback if the folio's mappability is NONE. If the folio is mapped, or if folio is not mapped but refcount is elevated for whatever reason, vcpu_run() fails and exits to userspace. If folio is not mapped and gmem holds the last refcount, set folio mappability to GUEST. Here's one issue I see based on the above understanding: Registration of the folio_put() callback only happens if the VMM actually tries to do vcpu_run(). For 4K folios I think this is okay since the 4K folio can be freed via the transition state K -> state I, but for hugetlb folios that have been split for sharing with userspace, not getting a folio_put() callback means never putting the hugetlb folio together. Hence, relying on vcpu_run() to add the folio_put() callback leaves a way that hugetlb pages can be removed from the system. I think we should try and find a path forward that works for both 4K and hugetlb folios. IIUC page._mapcount and page.page_type works as a union because page_type is only set for page types that are never mapped to userspace, like PGTY_slab, PGTY_offline, etc. Technically PGTY_guest_memfd is only set once the page can never be mapped to userspace, but PGTY_guest_memfd can only be set once mapcount reaches 0. Since mapcount is added in the faulting process, could gmem perhaps use some kind of .unmap/.unfault callback, so that gmem gets notified of all unmaps and will know for sure that the mapcount gets to 0? Alternatively, I took a look at the folio_is_zone_device() implementation, and page.flags is used to identify the page's type. IIUC a ZONE_DEVICE page also falls in the intersection of needing a folio_put() callback and can be mapped to userspace. Could we use a similar approach, using page.flags to identify a page as a guest_memfd page? That way we don't need to know when unmapping happens, and will always be able to get a folio_put() callback. [1] https://lpc.events/event/18/contributions/1758/attachments/1457/3699/Guestmemfd%20folio%20state%20page_type.pdf [2] https://android-kvm.googlesource.com/linux/+/764360863785ba16d974253a572c87abdd9fdf0b%5E%21/#F0 > This patch series doesn't necessarily impose all these transitions, > many of them would be a matter of policy. This just happens to be the > current way I've done it with pKVM/arm64. > > Cheers, > /fuad > > On Fri, 13 Dec 2024 at 16:48, Fuad Tabba <tabba@xxxxxxxxxx> wrote: >> >> This series adds restricted mmap() support to guest_memfd, as >> well as support for guest_memfd on arm64. It is based on Linux >> 6.13-rc2. Please refer to v3 for the context [1]. >> >> Main changes since v3: >> - Added a new folio type for guestmem, used to register a >> callback when a folio's reference count reaches 0 (Matthew >> Wilcox, DavidH) [2] >> - Introduce new mappability states for folios, where a folio can >> be mappable by the host and the guest, only the guest, or by no >> one (transient state) >> - Rebased on Linux 6.13-rc2 >> - Refactoring and tidying up >> >> Cheers, >> /fuad >> >> [1] https://lore.kernel.org/all/20241010085930.1546800-1-tabba@xxxxxxxxxx/ >> [2] https://lore.kernel.org/all/20241108162040.159038-1-tabba@xxxxxxxxxx/ >> >> Ackerley Tng (2): >> KVM: guest_memfd: Make guest mem use guest mem inodes instead of >> anonymous inodes >> KVM: guest_memfd: Track mappability within a struct kvm_gmem_private >> >> Fuad Tabba (12): >> mm: Consolidate freeing of typed folios on final folio_put() >> KVM: guest_memfd: Introduce kvm_gmem_get_pfn_locked(), which retains >> the folio lock >> KVM: guest_memfd: Folio mappability states and functions that manage >> their transition >> KVM: guest_memfd: Handle final folio_put() of guestmem pages >> KVM: guest_memfd: Allow host to mmap guest_memfd() pages when shared >> KVM: guest_memfd: Add guest_memfd support to >> kvm_(read|/write)_guest_page() >> KVM: guest_memfd: Add KVM capability to check if guest_memfd is host >> mappable >> KVM: guest_memfd: Add a guest_memfd() flag to initialize it as >> mappable >> KVM: guest_memfd: selftests: guest_memfd mmap() test when mapping is >> allowed >> KVM: arm64: Skip VMA checks for slots without userspace address >> KVM: arm64: Handle guest_memfd()-backed guest page faults >> KVM: arm64: Enable guest_memfd private memory when pKVM is enabled >> >> Documentation/virt/kvm/api.rst | 4 + >> arch/arm64/include/asm/kvm_host.h | 3 + >> arch/arm64/kvm/Kconfig | 1 + >> arch/arm64/kvm/mmu.c | 119 +++- >> include/linux/kvm_host.h | 75 +++ >> include/linux/page-flags.h | 22 + >> include/uapi/linux/kvm.h | 2 + >> include/uapi/linux/magic.h | 1 + >> mm/debug.c | 1 + >> mm/swap.c | 28 +- >> tools/testing/selftests/kvm/Makefile | 1 + >> .../testing/selftests/kvm/guest_memfd_test.c | 64 +- >> virt/kvm/Kconfig | 4 + >> virt/kvm/guest_memfd.c | 579 +++++++++++++++++- >> virt/kvm/kvm_main.c | 229 ++++++- >> 15 files changed, 1074 insertions(+), 59 deletions(-) >> >> >> base-commit: fac04efc5c793dccbd07e2d59af9f90b7fc0dca4 >> -- >> 2.47.1.613.gc27f4b7a9f-goog >>
